The Harness ’little powerhouse’ arrives: PilotDeck 2.0 goes open source with one-click setup and reliable delivery on long, hard tasks
Does the task always stall halfway through? Does adding one more document mean reworking the whole plan? Do you just want to change a single number or one bit of layout, yet the AI can only redo the entire file? And once the result is finally generated, do you still have to fix the formatting and page layout by hand? The real difficulty of AI at work was never producing the first draft — it is finishing the final deliverable together with a person.
On October 15, PilotDeck 2.0, an AI Agent workbench jointly developed by Tsinghua University's THUNLP, ModelBest, OpenBMB and AI9Stars, is officially open source. The "little powerhouse" family has a new member, and the Harness field has a new entrant.
Following its first release in May this year, PilotDeck 2.0 focuses its upgrade on complex tasks, especially real office scenarios. Word, Excel and PPT Skills have been fully optimized, connecting the entire flow of reading, organizing, calculating, generating, editing and checking, and delivering editable results directly. AgentLoop now supports stable resumption of long tasks, handling 2,000 interaction messages in extreme tests. Smart routing matches model combinations to each task, cutting social media task costs by more than 75% in measured tests and making the cost of continuous operation more controllable.
In the WorkbuddyBench office benchmark, PilotDeck sits in the same 80-point tier as Claude Code and Workbuddy, with comparable capability. When the base model is swapped for a lightweight 35B-A3B model, PilotDeck ranks first among the three with 79.4 points. This means that with better framework design, lightweight models can carry complex office tasks reliably too. From task processing to deliverable delivery, PilotDeck 2.0 takes AI out of the chat box and turns it into an intelligent work pod that can actually do the work and truly deliver.
This article includes video content. Please view it in the original WeChat post.
Light office work: end-to-end delivery for real office files
PilotDeck 2.0 distills document-processing experience into built-in Word, Excel and PPT Skills, supporting mainstream office formats including Word, Excel, PPT and PDF.
Following the human habit of "understand first, act next, check when done", once it receives a task it first understands the goal and the materials, then generates the file and can make targeted edits on request. When finished, it runs a multimodal check on document content and structure, spreadsheet data and formulas, and how the pages actually render. If it finds a problem, it goes back to editing and keeps fixing until the result meets the bar. It also keeps up with multiple rounds of adjustment and executes them accurately — like a dependable colleague who is always on call: understands what you mean, works fast, and changes exactly the right thing.
Each of the three Skills has its own specialty for different kinds of office work.
01 Word Skill: write, revise and check in one go
Given a template and project materials, it generates a complete document that follows the existing structure and layout, revises specified paragraphs on request, preserves tracked changes and comments, and handles pagination and page rendering as well.
02 Excel Skill: organize data, reconcile results
It reads and organizes spreadsheet data, handles formulas, computes results, recalculates when requirements change, and keeps an eye on how the sheet as a whole presents.
Take expense review as an example: just upload the approval forms and invoices and tell PilotDeck 2.0:
Review the reimbursement materials employees submitted, check whether each trip is a closed loop, and calculate the travel allowance.
PilotDeck 2.0 then reads the files one by one, links the time, itinerary and expense information in them, completes the checks and calculations, and puts together the review results.
03 PPT Skill: content and visual quality together
Describe what you need in words and PilotDeck 2.0 searches for materials on its own, producing the full content and visual edit of a deck in one stop, then checking the result through structure checks, before-and-after comparison and page-by-page rendering. Once the first draft is done, select the part you want to change and keep the conversation going — you can revise text or adjust the visual presentation without redoing the whole deck.
To verify what the Skills themselves contribute, the team ran an ablation: under identical conditions, enabling the built-in Skills raised PilotDeck's score from 78.30 to 79.40, a real gain for office tasks from the skill system.
More stable: long, hard tasks stay online at more controllable cost
The more complex the task, the worse it is when it drops halfway.
Office tasks especially tend to need repeated file reads, content processing and result checking, and can run for hours. A sudden network blip, a single overly long output, or the steady accumulation of context can knock a task offline at any moment, leaving the earlier work to start over from scratch.
PilotDeck 2.0 addresses this with an optimized AgentLoop that keeps state and continues execution when something unexpected happens. It also supports background running, so users do not have to sit in front of the computer: hand over the task and come back later to check the result.
This stability lets lightweight models take on complex tasks too. In tests, a 35B-A3B model kept executing hour-long tasks. PilotDeck 2.0 also further optimizes smart routing, letting different tasks flexibly match model combinations to balance quality and cost.
In social media task tests, enabling smart routing cut costs by more than 75%. Using a "strong primary model plus lightweight sub-models" combination, it reached 70.6 points at about one-sixth of the cost, above the 69.1 of a single flagship model.
So how far can PilotDeck 2.0 run stably on longer tasks with more steps?
In a 3D simulation case of the Russia-Ukraine war, PilotDeck 2.0 automatically completed open-source intelligence collection, cross-verification of content, 3D map construction, location marking and event presentation. The whole task accumulated 2,000 interaction messages and still ran normally and stably.
Easier to use: from install to follow-up edits, beginners get it instantly
PilotDeck 2.0 carries ease of use through the whole experience, from download and installation and first-time configuration to file upload, task adjustment and result review. The operations are clear and direct, and beginners can get up to speed quickly.
The official site offers desktop installers. After downloading, follow the prompts to pick a model provider, enter the endpoint and API key and test connectivity with one click, then choose a workspace and language to finish setup. Developers can also deploy with a single command or Docker Compose.
Inside the workbench you can create workspaces, check task progress and manage files. Upload several attachments at once; results can be previewed online and downloaded at any time. If you want to keep asking questions, just cite the file — and reuse it in later tasks.
Once a task has started you can still change direction, switch modes and models. Want to try a different approach? PilotDeck 2.0 can branch a new conversation from the current one while the original task keeps running, without disturbing each other. Conversation history is searchable, so finding a task or a file does not mean paging through logs.
PilotDeck 2.0 also supports common IM channels such as Feishu, WeChat, WeCom and QQ. Scan the code to create a bot, then send tasks directly in the chat window and receive progress updates and result pushes.
Open and transparent: local deployment, developer friendly
PilotDeck 2.0 is fully open source and transparent, and is not tied to any fixed model. Developers can choose models and deployment methods to fit their actual needs and build on it to extend Agent applications for their own business.
On data handling, PilotDeck 2.0 supports local and private models. Data transmission and processing are visible to users, who can keep internal or sensitive files on local machines or the intranet depending on the deployment plan. This gives companies sensitive about data leaving their premises a more privacy-protecting, self-controlled option, and provides a technical basis for deployments in classified environments that have passed security assessment and compliance review.
With an open, transparent architecture, local deployment capability and friendly support for secondary development, PilotDeck 2.0 meets differentiated needs across scenarios in a more flexible, controllable and secure way.
Lightweight in form, powerful in capability; flexible enough for diverse needs. In this intelligent work pod, users express what they need in natural language, the AI plans on its own and keeps pushing forward, and human-AI collaboration runs through the whole path from requirement to delivery.
Leave the complexity to the system and the convenience and control to the user. From people executing tasks around tools to systems organizing work around goals, PilotDeck 2.0 aims to become the foundation for the next generation of intelligent productivity, evolving how work is done and truly unlocking productivity.
➤ Open-source repository
🔗 https://github.com/OpenBMB/PilotDeck
➤ Website
🔗 https://pilotdeck.openbmb.cn
➤ Installers
Mac:
🔗 https://n00069-minicpm-5-0.obs.cn-north-4.myhuaweicloud.com/PilotDeck/PilotDeck-2026.1013.1-mac-arm64.dmg
Windows:
🔗 https://n00069-minicpm-5-0.obs.cn-north-4.myhuaweicloud.com/PilotDeck/PilotDeck-2026.1013.1-win-x64-setup.exe
➤ Latest installers
🔗 https://github.com/OpenBMB/PilotDeck/releases